Spatiotemporal Models for Environmental Pollution Monitoring

Project number: Research project P 12508 Project title: Spatio-Temporal Models for Environmental Pollution Monitoring ABSTRACT OF THE PROJECT The objective of this project is to build a model for environmental data with dynamic and spatial structure. The particular example under consideration are ozone data. An emphasis is put on problemoriented modeling. The aim is to find a suitable model class and appropriate identification procedures in an interative process using both apriori information (concerning meteorological and physical knowledge) and data from the Austrian ozone monitoring network. In modeling the spatio-temoral features of ozone concentrations, the understanding of the dynamics of ozone formation, ozone catabolism and ozone transport, dependent on meteorological input data (e.g. wind direction, wind speed), the topography and other air pollutants (e.g. Nox ), is essential. The objectives of modeling ozone are as follows: predication of ozone concentrations, spatial interpolation (estimation of ozone concentrations at non observed locations), monitoring of ozone concentrations and to obtain hints concerning the spatial allocation of monitoring stations (if sufficient time is left).

Final Report Project Objectives The objective of the project P-12508 MAT is to build models for environmental data with dynamic and spatial structure. The emphasis is on statistical models, i.e. models which are predominately black box and thus explain the data in purely input-output terms. The particular example under consideration is modeling and forecasting of half hourly ozone concentrations, daily ozone maxima and daily ozone profiles. Ozone and meteorological data from the Austrian ozone monitoring network (1993-1997) are used in the analysis. High concentrations of ozone are believed to be harmful to human health and to cause damage to crops. Thus, ozone is monitored routinely in many countries. In order to implement special warning and management strategies for ozone, forecasting models may be useful. Project Development In the first part of the project the time series of ozone sampled every half hour without considering the spatial structure is analyzed. This is done by appropriately modifying autoregressive models with exogenous variables (ARX) and using data from different monitoring sites in the eastern part of Austria. Model structures common to several monitoring sites are obtained and several statistical techniques to describe well known effects connected with ozone are analyzed. The prediction of the daily maximum of ozone (e.g. 6 hour average) is a major task of the project. Thus the second part of the project is centered on ARX models for the daily maximum of ozone using daily statistics of meteorological variables. To allow for time varying parameters also adaptive estimates using the Kalman filter are used. The emphasis of the third part of the project is the modeling of the diurnal profile of ozone using a factor model. The underlying idea is that diurnal ozone cycles share some common patterns. This naturally leads to the introduction of factor models, consisting of time invariant loadings (patterns) and dynamic factors. To allow a certain parameter variability in time, adaptive techniques using the Kalman filter are used. The fourth part is concerned with the spatio-temporal modeling of ozone. We consider multivariate time series models for monitoring stations as well as models which allow for spatial prediction at locations for which data are not available. Of special importance is the modeling of ozone maxima. Results obtained Simple structures are derived for all model classes described above. The diurnal cycle of half hourly data of ozone may be well described by ARX models with smoothly time varying coefficients . The diurnal pattern differs significantly between monitoring sites. A major question in this context is the spatio temporal variation of this diurnal pattern, i.e. if monitoring sites close to one another have similar model parameters at every time of the day. This is part of ongoing research and the results should give hints to develop a spatio temporal model. Adaptive models using the Kalman filter, which allow for time varying coefficients, show the best prediction performance for daily maxima. The mentioned models are reasonable for periods with slow variation in ozone. Extreme ozone events which are of particular importance for monitoring, however are up to now not satisfactory modeled. For the purpose of predicting the diurnal ozone pattern, we use site specific ARX models for the factors. This gives reasonable one day ahead predictions for the diurnal profile of ozone at a given monitoring site. Adaptive models using the Kalman filter improve the model performance. If prediction of daily ozone maxima is considered, the models give similar ``out of sample`` prediction errors if they are compared with ARX models for ozone maxima, i.e. the mentioned models are reasonable for periods with slow variation in ozone. Introducing spatio-temporal models such as spatio-temporal kriging does not satisfy the following purposes. First, the prediction performance for extreme ozone events is not improved and second, the models do not reconstruct the spatial field in a satisfying way. In our recent work we focus on spatial characteristics related to wind in order to better explain high ozone events and their place of occurrence. For the purpose of regional ozone forecasting the statistical models used show several shortcomings: (1) For monitoring and regulatory purposes, the regional ozone maximum is a key ozone summary. An underestimation of these peak ozone concentrations is common to various statistical forecasting models developed in the project. Lack of explanatory variables (e.g. regional wind fields, regional emission data etc.) and unknown nonlinearities within extreme ozone regimes may be blamed for this. (2) Statistical spatio-temporal estimation procedures produce smooth spatio-temporal ozone concentration fields. This may be sufficient for an ozone trend analysis or an comparison with outputs from deterministic models but from a regional perspective, this is often not satisfactory, since peak ozone values and the spatial location of their occurrence are badly forecasted. (3) The models used so far are predominately black box in form and thus explain the data in purely input-output terms. Since ozone production, depletion, diffusion and transportation is based on physical and chemical mechanisms, which are known in principle, there is a need for the implementation of a-priori knowledge and of "mechanistic interpretations". In our opinion, an important question is how to integrate this a-priori information in order to get better forecasts. Of special importance thereby is the use of this a-priori information in a parsimonious and efficient manner, i.e. to select only those parts of prior information which are of importance for the regional input output behavior. Summary Ozone and several meteorological variables are monitored at several sites in the eastern part of Austria. The times series structure of ozone data within this area have been analyzed by using autoregressive type models with an exogenous part as well as spatio-temporal models. The analysis gives insight into the dynamics of ozone and its explanation by meteorological variables. Extreme ozone events are badly explained within the proposed framework.